Risk Measures for Skew Normal Mixtures
Mauro Bernardi MEMOTEF Department Sapienza University of Rome
July 4, 2012
Abstract
Finite mixtures of Skew distributions have become increasingly popular in the last few years as a flexible tool for handling data displaying several different characteristics such as multimodality, asymmetry and fat-tails. Examples of such data can be found in financial and actuarial applications as well as biological and epidemiological analysis. In this paper we will show that a convex linear combination of multivariate Skew Normal mixtures can be represented as finite mixtures of univariate Skew Normal distributions. This result can be useful in modeling portfolio returns where the evaluation of extremal events is of great interest. We provide analytical formula for different risk measures like the Value-at-Risk and the Expected Shortfall probability.
Keywords: Finite mixtures, Skew Normal distributions, Value-at-Risk, Expected Shortfall probability.
1 Introduction
The implementation of finite mixture has become increasingly popular in many disciplines, such as biological sequences analysis, econometrics, machine learning, actuarial science, finance and epidemiology. Often data display strong asymmetry, fat tails and multimodality features that are usually shared by different subpopulations. In this context mixtures of asymmetric distributions have been adopted in different areas, see for example Fr¨uhwirth-Schnatter and Pyne [6], Bernardi et al. [4] and Haas and Mittnik [8]. Among the skewed distrubutions, the Skew Normal of Azzalini [1]
and the Skew Student-t of Azzalini and Capitanio [3] have become widely employed because of the attractive properties they share with their symmetric counterparts, and the greater shape flexibility introduced by the additional asymmetry parameter.
In what follows we will consider finite mixture of multivariate Skew Normal distributions and linear combinations of them. We will show that a linear combination of mixtures of multivariate Skew Normals admits a closed form representation as a finite mixtures of univariate Skew Normals.
This result can be very useful for example when we model the distribution of the portfolio return, defined as a convex linear combination of the portfolio assets assumed to be a mixture of multivariate Skew Normals. In fact, it is well known in the financial literature that financial returns strongly deviates from the Gaussian assumption.
Moreover, since investors holding short or long positions on portfolios of risky assets are mainly interested in evaluating the risk associated to their portfolios, it is important to calculate some risk measures like the Value-at-Risk, (VaR), and the Expected Shortfall probability, (ES). Usually those risk measure are estimated by historical or Monte Carlo simulations. In this paper, under the assumption of a joint multivariate Skew Normal mixture for the assets, we will provide an analytical closed form expression for the VaR, the ES and related measures, as functions of the model parameters.
The paper is organized as follows: in Section 2 we prove that the multivariate Skew Normal mixtures are closed with respect to linear combinations. In Section 3 we provide analytical formulas for the Value-at-Risk, the Expected Shortfall probability, and related risk measures both for Skew Normals and mixtures of them. Few remarks and possible extensions are discussed in Section 4.
2 Linear combinations of multivariate Skew Normal mixtures
Finite mixture of multivariate Skew Normal distributions (see Azzalini and Dalla Valle [2]) for the d-dimensional observation data y= (y1,y2, . . . ,yd) can be defined as
π(y|ξ,Ω,α,η) =
L
X
l=1
ηlfMSN(y|ξl,Ωl,αl), (2.1) where fMSN(y|ξ,Ω,α) is the density of the multivariate Skew Normal distribution defined in Azzalini and Dalla Valle [2]:
fMSN(y|ξ,Ω,α) = 2Φ γTΩ−1(y−ξ) p1−γTΩ−1γ
! 1 (2π)d2 |Ω|12
exp
−1
2(y−ξ)TΩ−1(y−ξ)
, (2.2) where ξ ∈ Rd is a d−dimensional vector of location parameters, Ω is a positive definite square matrix of dimension d, and γ is defined as a reparameterization of the d−dimensional vector of skewness parameters α, in the following way
γ = Ω12α
√1 +αTα =Ω12δ. (2.3)
The parameters ηl, l= 1,2, . . . , L are the component weights satisfying 0≤ηl≤1,∀l= 1,2, . . . , L, and PL
l=1ηl= 1. We denote with φ(x) and Φ (x) the density function and the cumulative density function of a scalar Gaussian distribution, and with δ = √α
1+α2, the univariate counterpart of the δ parameter defined in the previous equation (2.3). The univariate Skew Normal distribution can be obtained from the previous definition by setting the dimension dof the vector of observation y equal to 1.
In what follows we prove our main result contained in Theorem 2.1 showing that the distribution of linear combinations of multivariate Skew Normal mixtures is a mixture of univariate Skew Normals.
Before stating the theorem it is useful to recall the following known results providing the Moment Generating Function (MGF) of univariate and multivariate Skew Normal distributions and their extension to mixtures of them.
Corollary 2.1. Let X ∼ SN(ξ, ω, α) be a univariate Skew Normal distribution, and X ∼ MSN (ξ,Ω,α) be a d-dimensional Skew Normal distribution, then the MGF of X and X are respectively
MX(t) = 2 exp
ξt+t2ω2 2
Φ (δωt) (2.4)
MX(t) = 2 exp
tTξ+tTΩt 2
Φ
δTΩ12t
. (2.5)
Proof. See Genton [7], page 7 and 17. 22
Corollary 2.2. Let Y ∼ PL
l=1ηlfSN(ξl, ωl, αl), be a univariate Skew Normal mixture, and Y ∼ PL
l=1ηlfMSN(y|ξl,Ωl,αl) be ad-dimensional Skew Normal mixture, then the MGF of Yand Y are respectively
MY(t) =
L
X
l=1
ηl
2 exp
ξlt+ t2ω2l 2
Φ (δlωlt)
(2.6)
MY(t) =
L
X
l=1
ηl
2 exp
tTξl+tTΩlt 2
Φ
δTlΩ
1
l2t
. (2.7)
Proof. The proof is straightforward. 22
Theorem 2.1 (Linear combinations of multivariate Skew Normal mixtures). Let Y ∼ PL
l=1ηlfMSN(y|ξl,Ωl,αl), assume w ∈ Rd is d-dimensional vector of real coefficients, such that Pd
j=1wj = 1, then
Z=wTY (2.8)
has density function
fZ(z) =
L
X
l=1
ηlfSN
z|ξ˜l,ω˜l,α˜l ,
where ξ˜k=wTξl, ω˜l = wTΩlw1
2, δ˜l =wTδl, for l= 1,2, . . . , L. The shape parameters α˜l can be recovered by inverting the relation δ˜l= √α˜l
1+ ˜α2l, getting α˜l= qδ˜l
1−˜δ2l
,∀l= 1,2, . . . , L.
Proof. The MGF of the (scalar) random variable Z=wTY is equal to MZ(t) = E
h etZ
i
=E h
et(wTY)i
= Z
Rd
expn t
wTy o×
L
X
l=1
ηl2Φ
γTlΩ−l 1(x−ξl) q
1−γTlΩ−l 1γl
1
(2π)d2|Ωl|12 exp
−1
2(y−ξl)TΩ−l 1(y−ξl)
dy
After some calculations the MGF can be written as
MZ(t) =
L
X
l=1
ηl2 exp
t
wTξl
+t2wTΩlw 2
Z
Rd
Φ
γTlΩ−l 1(x−ξl) q
1−γTlΩ−l 1γl
1
(2π)d2 |Ωl|12 × exp
(
−1 2
y−ξl−Ω
1
l2wt T
Ω−l 1
y−ξl−Ω
1
l2wt )
dy.
Now, considering the transformation z=Ω−
1
l 2
y−ξl−Ω
1
l2wt
, we get
MY(t) =
L
X
l=1
ηl2 exp
t wTξl
+t2wTΩlw 2
Z
Rd
Φ
γTlΩ−
1
l 2z+γTlΩ−12wt q
1−γTl Ω−l1γl
exp
−12zTz (2π)d2
dz.
By applying the result presented in Ellison [5], detailed in appendix, we obtain
MY(t) =
L
X
l=1
ηl2 exp
t
wTξl
+t2wTΩlw 2
Φ
δTlwt
,
which corresponds to the MGF of a univariate mixture of Skew Normal distribution having the following parameters: n
ηl,ξ˜l=wTξl,ω˜l= wTΩkw1
2,˜δl=δTlw oL
l=1. 22
As a byproduct of this result we have that setting w = (0, . . . ,1, . . . ,0) we obtain the marginal distribution of thei-th component of the vector Y.
3 Risk Measures
Investors holding short or long positions on portfolios of risky assets are mainly interested in evaluating the exposure to risk associated to their global positions. Assuming that the distribution of the portfolio returns follow a multivariate mixture of Skew Normals we calculate different risk measures for the resulting linear combination of them. The results stated in previous Section 2 provide the theoretical framework to calculate such risk measures. As risk measures, we consider, the probability of shortfall (PS), i.e. the probability that a portfolio falls short some target level, the probability of outperformance (PO), i.e the probability of outperforming the target level, and the more familiar measures known as target shortfall (TS) and tail conditional expectation (TCE).
All these measures can be interpreted as a partial moments of a given order of the corresponding random variable. They are evaluated for univariate Skew Normal distributions in Section 3.1 and subsequently extended to mixtures of Skew Normals in Section 3.2. As a latter result we show how to evaluate the Value-at-Risk and the Expected Shortfall probability using the provided measures.
3.1 Risk measures for univariate Skew Normal distributions
In this Section we provide explicit formulas to evaluate risk measures of a univariate Skew Normal random variable.
Theorem 3.1 (PS and PO for univariate Skew Normal). LetY∼ SN(ξ, ω, α), the probability of shortfall PS, i.e. the proability that the random variable Y falls short of the target yp is
PSY(yp, ξ, ω, α) =FX
yp−ξ ω , α
, (3.1)
while the probability that the random variable Y outperforms the target yp, i.e. the probability of outperformance PO, is
POY(yp, ξ, ω, α) = 1−FX
yp−ξ ω , α
= 1−PSY(yp, ξ, ω, α), (3.2)
where FX(x, α) is the cdf of a standardized Skew Normal distribution, i.e. X∼ SN(α) evaluated at x.
Proof. The shortall probability, PS, is the probability that a univariate random variable, Y ∼ SN(ξ, ω, α), falls short the threshold yp, and can be evaluated as the zero-th order lower partial moment (LPM) of the random variable Y with respect to the threshold yp ∈R
PSY(yp, ξ, ω, α) = LPMY(yp,0)
= E
n
[yp−Y]0+o
=
Z yp−ξ
ω
−∞
2Φ (αx)φ(x)dx
= FX
yp−ξ ω , α
.
The probability of outperformance, PO, is the probability that a univariate random variable, Y ∼ SN(ξ, ω, α), outperforms the threshold yp, and can be evaluated as the zero-th order upper partial moment (UPM) of the random variable Y with respect to the threshold yp ∈R
POY(yp, ξ, ω, α) = UPMY(yp,0)
= E
n
[Y−yp]0+o
=
Z +∞
yp−ξ ω
2Φ (αx)φ(x)dx
= 1−PSY(yp, ξ, ω, α).
22
Theorem 3.2 (TS for univariate Skew Normal). Let Y∼ SN(ξ, ω, α), then the target shortfall TS, of Y is
TSY(yp, ξ, ω, α) = (ξ−yp)POY(yp, ξ, ω, α) +ω[2Φ (αxp)φ(xp) +δb(1−Φ (zp))], (3.3) where b=
q2
π, zp=√
1 +α2xp and xp = ypω−ξ.
Proof. The target shortfall, TS, can be evaluated as the first order upper partial moment UPM of the random variable Y∼ SN(ξ, ω, α), with respect to the threshold yp ∈R
TSY(yp, ξ, ω, α) = UPMY(yp,1)
≡ En
[Y−yp]1+o
=
Z +∞
yp−ξ ω
2Φ (αx)φ(x) dx.
After some algebra we get:
TSY(yp, ξ, ω, α) = (ξ−yp) Z +∞
yp−ξ ω
2Φ (αx)φ(x) dx+ω Z +∞
yp−ξ ω
2xΦ (αx)φ(x) dx
= (ξ−yp)POY(yp, ξ, ω, α) +ω Z +∞
yp−ξ ω
2xΦ (αx)φ(x) dx.
Integrating by parts we obtain Z +∞
yp−ξ ω
2xΦ (αx)φ(x) dx = h
−2φ(x) Φ (αx)i+∞ xp
+αb Z +∞
xp
exp
−x2 2
φ(αx) dx
= 2φ(xp) Φ (αxp) +αb Z +∞
xp
√1 2π exp
(
− 1 +α2 x2 2
) dx
= 2φ(xp) Φ (αxp) + αb
√1 +α2 Z +∞
zp
φ(z) dz
= 2φ(xp) Φ (αxp) +δb[1−Φ (zp)]
where b, xp, zp and δ are defined as before. 22
Finally, in the following theorem we use previous results to calculate the tail conditional expecation for a univariate Skew Normal distributions.
Theorem 3.3 (TCE for univariate Skew Normal). Let Y ∼ SN(ξ, ω, α), then the tail conditional expectation of Y, i.e. the mean of Y truncated below the threshold yp, is
TCEY(yp, ξ, ω, α) =ξ+ ωb PSY(yp, ξ, ω, α)
hδΦ (zp)−√
2πφ(xp) Φ (αxp)i
. (3.4)
where b = q2
π and PSY(x, ξ, ω, α) denotes the probability of shortfall of a Skew Normal r.v.
evaluated at x, defined in equation (3.1).
Proof. Let X∼ SN(α) be a standardized Skew Normal density, consider Y=ξ+ωX, thenY has a Skew Normal distribution Y∼ SN(ξ, ω, α), the TCE ofY is
TCEY(yp, ξ, ω, α) = E(Y|Y≤yp)
= ξ+ωE
X|X≤ yp−ξ ω
= ξ+ωE(X|X≤xp)
= ξ+ωTCEX(xp, α), (3.5)
where xp = ypω−ξ, and TCEX(xp, α) is the TCE of a standardized Skew Normal distribution, X∼ SN(α) and can be evaluated as follows
TCEX(xp, α) = E(X|X≤xp)
= 1
FX(xp, α) Z xp
−∞
xfX(x, α) dx (3.6)
where fSN(x, α) = 2φ(x) Φ (αx), is the probability density function of the standardized Skew Normal distribution and FX(xp, α) is the corresponding cdf. By substituting this last expression for the density in the previous defintion of tail conditional expectation (3.6), it becomes
TCEX(xp, α) = b FX(xp, α)
Z xp
−∞
xexp
−x2 2
Φ (αx) dx, where b=
q2
π. Integrating by parts, we have
TCEX(xp, α) = b FX(xp, α)
−exp
−x2 2
Φ (αx)
xp
−∞
+α Z xp
−∞
exp
−x2 2
φ(αx) dx
= b
FX(xp, α)
"
−exp (
−x2p 2
)
Φ (αxp) +α Z xp
−∞
exp
−x2 2
φ(αx) dx
#
, (3.7) where the last integral in the previous formula (3.7) can be evaluated as follows
Z xp
−∞
exp
−x2 2
φ(αx) dx = Z xp
−∞
√1 2π exp
(
− 1 +α2 x2 2
) dx
= 1
√1 +α2 Z zp
−∞
√1 2π exp
−z2 2
dz
= Φ (zp)
√1 +α2, where zp = √
1 +α2xp. By substituting this last expression into the previous equation (3.7) we obtain the final form for the TCE
TCEX(xp, α) = b FX(xp, α)
α
√1 +α2Φ (zp)−√
2πφ(xp) Φ (αxp)
= b
FX(xp, α) h
δΦ (zp)−√
2πφ(xp) Φ (αxp)i and, by exploiting equation (3.5), we obtain
TCEY(yp, ξ, ω, α) =ξ+ ωb PSY(yp, ξ, ω, α)
hδΦ (zp)−√
2πφ(xp) Φ (αxp)i ,
where zp=√
1 +α2xp, andxp = ypω−ξ. 22
3.2 Risk measures for univariate Skew Normal mixtures
In this Section we extend the risk measures presented in the previous section to the mixtures of univariate Skew Normal distributions.
Theorem 3.4 (PS, PO and TS for mixtures of Skew Normals). Let Y ∼ PL
l=1ηlfSN(y|ξl, ωl, αl), then the shortfall probability PS, the probability of outperformance, PO, and the target shortfall TS of Y are convex linear combinations of the PS, PO and TS, of the component densities evaluated as in equations (3.1), (3.2) and (3.3):
PSY(yp, L) =
L
X
l=1
ηlPSl(yp, ξl, ωl, αl) (3.8)
POY(yp, L) =
L
X
l=1
ηlPOl(yp, ξl, ωl, αl) (3.9)
TSY(yp, L) =
L
X
l=1
ηlTSl(yp, ξl, ωl, αl). (3.10)
Proof. The proof is straightforward. 22
Theorem 3.5 (TCE for Skew Normal mixtures). Let Y ∼ PL
l=1ηlfSN(y|ξl, ωl, αl), then the tail conditional expectation of Y is a convex linear combination of the tail conditional expectations of the components:
TCEY(yp, L) =
L
X
l=1
πlTCEl(yp, ξl, ωl, αl) (3.11) where the weights are πl=ηlPSlPS(yp,ξl,ωl,αl)
Y(yp,L) , l= 1,2, . . . , L, with PL
l=1πl= 1.
Proof.
TCEY(yp, L) = E(Y|Y≤yp)
= 1
PSY(yp, L) Z yp
−∞
y
" L X
l=1
ηlf(y|ξl, ωl, αl)
# dy
= 1
PSY(yp, L)
L
X
l=1
ηl Z yp
−∞
yf(y|ξl, ωl, αl)dy
=
L
X
l=1
ηlPSl(yp, ξl, ωl, αl)
PSY(yp, L) TCEl(yp, ξl, ωl, αl)
=
L
X
l=1
πlTCEl(yp, ξl, ωl, αl), where πl = ηlPSPSl(yp,ξl,ωl,αl)
Y(yp,L) , with PL
l=1πl = 1, and TCEl(yp, ξl, ωl, αl),∀l = 1,2, . . . , L, can be
evaluated using equation (3.4). 22
We now briefly discuss the evaluation of the well known measures of risk: the Value-at-Risk (VaR), and the Expected Shortfall Probability (ES).
The VaR of portfolio return Z =wTY at confidence level λ is defined as the smallest number z0 such that the probability that Zfalls short z0 is not larger than 1−λ
VaRλ(Z) = inf{z0∋P(Z<z0)≤1−λ}
= inf{z0∋F(z0)≤1−λ}
= FZ−1(1−λ)
where FZ() is the cdf of Z, FZ−1() is the inverse function of FZ() provided one exists, and the last inequality holds for continuous distributions. Under the assumption that the risky assets Y have a mixture of multivariate Skew Normal distributions we have shown in Theorem 2.1 that the portfolio returnZ=wTY is distributed as a univariate mixture of Skew Normals. Hence, the VaR of the portfolio returnZ=wTYat fixedλconfidence level is evaluated as the unique solution with respect to z of the following equation:
PSZ(z, L)−(1−λ) = 0, (3.12)
where PSZ(z, L) is the probability of shortfall of the distribution ofZ.
The Expected Shortfall probability of the portfolio return Zis defined as ESλ(Z) = E
Z|Z≤VaRλ(Z)
= 1
1−λ Z
Z≤VaRλ(Z)
zf(z) dz (3.13)
that is, the λlevel expected shortfall is the average loss smaller or equal than the λ level quantile loss VaRλ(Z); this value is analytically tractable for mixture of Skew Normals, and coincides with the TCE evaluated at the VaRλ(Z), i.e.
TCEZ(VaRλ(Z), L) =
L
X
l=1
πlTCEl(VaRλ(Z), ξl, ωl, αl), (3.14)
where TCEZ(VaRλ(Z), L) is evaluated as in equation (3.11).
4 Conclusion
In this paper we provide a new result for linear combinations of multivariate Skew Normal distributions. This result can be useful when we deal with portfolio returns as convex linear combinations of different risky assets, modeled by mixtures of multivariate Skew Normal distributions. Since investors are interested in evaluating the risk associated to their global portfolios, we calculate risk measures as Value-at-Risk and Expected Shortfall Probability. The provided formulas can be useful in applications for portfolio optimization.
Appendix
In this appendix we state the main result contained in Ellison [5] which is useful in deriving the moment generating function of the Skew Normal distribution and for proving Theorem 2.1.
Theorem 4.1. LetXa Gaussian random variable with meanµand varianceσ2, i.e. X∼ N µ, σ2 , then
E(Φ (X)) = Φ
µ
√1 +σ2
. (4.1)
For the proof of the theorem we refer the reader to the work of Ellison [5], and in particular to the first corollary to the Theorem 2.
Acknowledgments This work has been partially supported by the 2010 and 2011 Sapienza University of Rome Research Project.
References
[1] Azzalini, A., (1985). A class of distributions which includes the normal ones, Scandinavian Journal of Statistics, Vol. 12, pp. 171-178.
[2] Azzalini, A. and Dalla Valle A., (1996). The multivariate Skew Normal distribution.
Biometrika, Vol. 83, pp. 715-726.
[3] Azzalini, A. and Capitanio, A., (1999). Statistical application of multivariate Skew Normal distribution. Journal of the Royal Statistical Society, Series B, Vol. 61, 579–602.
[4] Bernardi M., Maruotti, A., and Petrella, L., (2012). Skew Mixture Models for Loss Distributions: A Bayesian Approach. Woring Paper, MEMOTEF Detp, Sapienza University of Rome.
[5] Ellison, B. E., (1964). Two Theorems for Inferences about the Normal Distribution with Applications in Acceptance Sampling. Journal of the American Statistical Association, Vol.
59, pp. 89–95.
[6] Fr¨uhwirth-Schnatter, S. and Pyne, S., (2010). Bayesian inference for finite mixtures of univariate and multivariate Skew-Normal and Skew-t distributions. Biostatistics, Vol. 11, 317–36.
[7] Genton, M. G., (2004). Skew-Elliptical distriutions and their applications: a journey beyond normality. Chapman and Hall.
[8] Haas, M. and Mittnik, S., (2009). Skew Normal mixture and Markov switching GARCH processes. Studies in Nonlinear Dynamics Econometrics, Vol. 14, Article 1.